Convert a few lit tests that use "lc2i | llvm-mc | llvm-objdump" to use p2i.

Convert a few lit tests that use "lc2i | llvm-mc | llvm-objdump" to use p2i.

Makes it more uniform for a later bulk refactoring of
"p2i | llvm-mc | llvm-objdump", so it only needs to match p2i
and not lc2i. Some of the test expectations change though.

Also there was a stray XCHECK that I don't remember why
it was XCHECK... change that to CHECK.

BUG= none
R=stichnot@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=7b300674d7d83ac5994520a0da7cf60a0ac07f4f

[jvoung (off chromium), 2015-02-13 05:10:55]

jvoung@chromium.org changed reviewers:
+ kschimpf@google.com, stichnot@chromium.org

[jvoung (off chromium), 2015-02-13 05:13:14]

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
File tests_lit/llvm2ice_tests/phi.ll (left):

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
tests_lit/llvm2ice_tests/phi.ll:95: ; CHECK: mov {{.*}}[[ADDR:.*1000]]
This used to form a more complex addressing mode [x + 1000]. Not sure why it
doesn't anymore.

Interestingly, the new code is more compact, since the "1000" only has to be
encoded once instead of at both the load and the store.

[Jim Stichnoth, 2015-02-13 14:16:12]

otherwise lgtm

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
File tests_lit/llvm2ice_tests/phi.ll (left):

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
tests_lit/llvm2ice_tests/phi.ll:95: ; CHECK: mov {{.*}}[[ADDR:.*1000]]
On 2015/02/13 05:13:14, jvoung wrote:
> This used to form a more complex addressing mode [x + 1000]. Not sure why it
> doesn't anymore.
> 
> Interestingly, the new code is more compact, since the "1000" only has to be
> encoded once instead of at both the load and the store.

This is probably because p2i removes %__4 and %__6 along with the inttoptr
instructions, and then address mode inference stops before it reaches %interior
because %interior is live across multiple basic blocks (a heuristic to manage
register pressure).

Try the following instead.  It forces sort of a fake copy from %interior to keep
consistent behavior between p2i and lc2i.  I think it's important to keep the
"1000" marker in the test, so that non register-allocated temporaries don't
accidentally match the pattern.

diff --git a/tests_lit/llvm2ice_tests/phi.ll b/tests_lit/llvm2ice_tests/phi.ll
index 0099fed..e1257a7 100644
--- a/tests_lit/llvm2ice_tests/phi.ll
+++ b/tests_lit/llvm2ice_tests/phi.ll
@@ -2,9 +2,7 @@
 ; it tests that it does the right thing when it tries to enable
 ; compare/branch fusing.
 
-; TODO(kschimpf) Find out why lc2i must be used.
-; REQUIRES: allow_llvm_ir_as_input
-; RUN: %lc2i -i %s --args -O2 --verbose none --phi-edge-split=0 \
+; RUN: %p2i -i %s --args -O2 --verbose none --phi-edge-split=0 \
 ; RUN:   | llvm-mc -triple=i686-none-nacl -filetype=obj \
 ; RUN:   | llvm-objdump -d -symbolize -x86-asm-syntax=intel - | FileCheck %s
 
@@ -58,12 +56,14 @@ entry:
 body:
   %merge = phi i32 [ %arg, %entry ], [ %elt, %body ]
   %interior = add i32 %merge, 1000
-  %__4 = inttoptr i32 %interior to i32*
+  %interior__4 = add i32 %interior, 0
+  %__4 = inttoptr i32 %interior__4 to i32*
   %elt = load i32* %__4, align 1
   %cmp = icmp eq i32 %elt, 0
   br i1 %cmp, label %exit, label %body
 exit:
-  %__6 = inttoptr i32 %interior to i32*
+  %interior__6 = add i32 %interior, 0
+  %__6 = inttoptr i32 %interior__6 to i32*
   store i32 %arg, i32* %__6, align 1
   ret i32 %arg
 }

[jvoung (off chromium), 2015-02-13 16:37:52]

New patchsets have been uploaded after l-g-t-m from stichnot@chromium.org

[jvoung (off chromium), 2015-02-13 16:40:08]

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
File tests_lit/llvm2ice_tests/phi.ll (left):

https://codereview.chromium.org/927433002/diff/20001/tests_lit/llvm2ice_tests...
tests_lit/llvm2ice_tests/phi.ll:95: ; CHECK: mov {{.*}}[[ADDR:.*1000]]
On 2015/02/13 14:16:12, stichnot wrote:
> On 2015/02/13 05:13:14, jvoung wrote:
> > This used to form a more complex addressing mode [x + 1000]. Not sure why it
> > doesn't anymore.
> > 
> > Interestingly, the new code is more compact, since the "1000" only has to be
> > encoded once instead of at both the load and the store.
> 
> This is probably because p2i removes %__4 and %__6 along with the inttoptr
> instructions, and then address mode inference stops before it reaches
%interior
> because %interior is live across multiple basic blocks (a heuristic to manage
> register pressure).
> 
> Try the following instead.  It forces sort of a fake copy from %interior to
keep
> consistent behavior between p2i and lc2i.  I think it's important to keep the
> "1000" marker in the test, so that non register-allocated temporaries don't
> accidentally match the pattern.
> 
> diff --git a/tests_lit/llvm2ice_tests/phi.ll b/tests_lit/llvm2ice_tests/phi.ll
> index 0099fed..e1257a7 100644
> --- a/tests_lit/llvm2ice_tests/phi.ll
> +++ b/tests_lit/llvm2ice_tests/phi.ll
> @@ -2,9 +2,7 @@
>  ; it tests that it does the right thing when it tries to enable
>  ; compare/branch fusing.
>  
> -; TODO(kschimpf) Find out why lc2i must be used.
> -; REQUIRES: allow_llvm_ir_as_input
> -; RUN: %lc2i -i %s --args -O2 --verbose none --phi-edge-split=0 \
> +; RUN: %p2i -i %s --args -O2 --verbose none --phi-edge-split=0 \
>  ; RUN:   | llvm-mc -triple=i686-none-nacl -filetype=obj \
>  ; RUN:   | llvm-objdump -d -symbolize -x86-asm-syntax=intel - | FileCheck %s
>  
> @@ -58,12 +56,14 @@ entry:
>  body:
>    %merge = phi i32 [ %arg, %entry ], [ %elt, %body ]
>    %interior = add i32 %merge, 1000
> -  %__4 = inttoptr i32 %interior to i32*
> +  %interior__4 = add i32 %interior, 0
> +  %__4 = inttoptr i32 %interior__4 to i32*
>    %elt = load i32* %__4, align 1
>    %cmp = icmp eq i32 %elt, 0
>    br i1 %cmp, label %exit, label %body
>  exit:
> -  %__6 = inttoptr i32 %interior to i32*
> +  %interior__6 = add i32 %interior, 0
> +  %__6 = inttoptr i32 %interior__6 to i32*
>    store i32 %arg, i32* %__6, align 1
>    ret i32 %arg
>  }

Ah interesting -- make sense. Done.

[jvoung (off chromium), 2015-02-13 17:47:14]

Committed patchset #4 (id:60001) manually as
7b300674d7d83ac5994520a0da7cf60a0ac07f4f (presubmit successful).